src/runtime/thread.c

   1 /*
   2  * This software is part of the SBCL system. See the README file for
   3  * more information.
   4  *
   5  * This software is derived from the CMU CL system, which was
   6  * written at Carnegie Mellon University and released into the
   7  * public domain. The software is in the public domain and is
   8  * provided with absolutely no warranty. See the COPYING and CREDITS
   9  * files for more information.
  10  */
  11
  12 #include "sbcl.h"
  13
  14 #include <stdlib.h>
  15 #include <stdio.h>
  16 #include <string.h>
  17 #ifndef LISP_FEATURE_WIN32
  18 #include <sched.h>
  19 #endif
  20 #include <signal.h>
  21 #include <stddef.h>
  22 #include <errno.h>
  23 #include <sys/types.h>
  24 #ifndef LISP_FEATURE_WIN32
  25 #include <sys/wait.h>
  26 #endif
  27
  28 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
  29 #include <mach/mach.h>
  30 #include <mach/mach_error.h>
  31 #include <mach/mach_types.h>
  32 #endif
  33
  34 #include "runtime.h"
  35 #include "validate.h"           /* for CONTROL_STACK_SIZE etc */
  36 #include "alloc.h"
  37 #include "thread.h"
  38 #include "arch.h"
  39 #include "target-arch-os.h"
  40 #include "os.h"
  41 #include "globals.h"
  42 #include "dynbind.h"
  43 #include "genesis/cons.h"
  44 #include "genesis/fdefn.h"
  45 #include "interr.h"             /* for lose() */
  46 #include "gc-internal.h"
  47
  48 #ifdef LISP_FEATURE_WIN32
  49 /*
  50  * Win32 doesn't have SIGSTKSZ, and we're not switching stacks anyway,
  51  * so define it arbitrarily
  52  */
  53 #define SIGSTKSZ 1024
  54 #endif
  55
  56 #if defined(LISP_FEATURE_DARWIN) && defined(LISP_FEATURE_SB_THREAD)
  57 #define QUEUE_FREEABLE_THREAD_STACKS
  58 #endif
  59
  60 #ifdef LISP_FEATURE_FREEBSD
  61 #define CREATE_CLEANUP_THREAD
  62 #define LOCK_CREATE_THREAD
  63 #endif
  64
  65 #define ALIEN_STACK_SIZE (1*1024*1024) /* 1Mb size chosen at random */
  66
  67 struct freeable_stack {
  68 #ifdef QUEUE_FREEABLE_THREAD_STACKS
  69     struct freeable_stack *next;
  70 #endif
  71     os_thread_t os_thread;
  72     os_vm_address_t stack;
  73 };
  74
  75
  76 #ifdef QUEUE_FREEABLE_THREAD_STACKS
  77 static struct freeable_stack * volatile freeable_stack_queue = 0;
  78 static int freeable_stack_count = 0;
  79 pthread_mutex_t freeable_stack_lock = PTHREAD_MUTEX_INITIALIZER;
  80 #else
  81 static struct freeable_stack * volatile freeable_stack = 0;
  82 #endif
  83
  84 int dynamic_values_bytes=4096*sizeof(lispobj);  /* same for all threads */
  85 struct thread * volatile all_threads;
  86 extern struct interrupt_data * global_interrupt_data;
  87
  88 #ifdef LISP_FEATURE_SB_THREAD
  89 pthread_mutex_t all_threads_lock = PTHREAD_MUTEX_INITIALIZER;
  90 #ifdef LOCK_CREATE_THREAD
  91 static pthread_mutex_t create_thread_lock = PTHREAD_MUTEX_INITIALIZER;
  92 #endif
  93 #endif
  94
  95 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
  96 extern lispobj call_into_lisp_first_time(lispobj fun, lispobj *args, int nargs);
  97 #endif
  98
  99 static void
 100 link_thread(struct thread *th)
 101 {
 102     if (all_threads) all_threads->prev=th;
 103     th->next=all_threads;
 104     th->prev=0;
 105     all_threads=th;
 106 }
 107
 108 #ifdef LISP_FEATURE_SB_THREAD
 109 static void
 110 unlink_thread(struct thread *th)
 111 {
 112     if (th->prev)
 113         th->prev->next = th->next;
 114     else
 115         all_threads = th->next;
 116     if (th->next)
 117         th->next->prev = th->prev;
 118 }
 119 #endif
 120
 121 static int
 122 initial_thread_trampoline(struct thread *th)
 123 {
 124     lispobj function;
 125 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 126     lispobj *args = NULL;
 127 #endif
 128     function = th->no_tls_value_marker;
 129     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 130     if(arch_os_thread_init(th)==0) return 1;
 131     link_thread(th);
 132     th->os_thread=thread_self();
 133 #ifndef LISP_FEATURE_WIN32
 134     protect_control_stack_guard_page(1);
 135 #endif
 136
 137 #if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
 138     return call_into_lisp_first_time(function,args,0);
 139 #else
 140     return funcall0(function);
 141 #endif
 142 }
 143
 144 #define THREAD_STRUCT_SIZE (THREAD_CONTROL_STACK_SIZE + BINDING_STACK_SIZE + \
 145                             ALIEN_STACK_SIZE + dynamic_values_bytes + \
 146                             32 * SIGSTKSZ)
 147
 148 #ifdef LISP_FEATURE_SB_THREAD
 149
 150 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 151
 152 static void
 153 queue_freeable_thread_stack(struct thread *thread_to_be_cleaned_up)
 154 {
 155      if (thread_to_be_cleaned_up) {
 156         pthread_mutex_lock(&freeable_stack_lock);
 157         if (freeable_stack_queue) {
 158             struct freeable_stack *new_freeable_stack = 0, *next;
 159             next = freeable_stack_queue;
 160             while (next->next) {
 161                 next = next->next;
 162             }
 163             new_freeable_stack = (struct freeable_stack *)
 164                 os_validate(0, sizeof(struct freeable_stack));
 165             new_freeable_stack->next = NULL;
 166             new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 167             new_freeable_stack->stack = (os_vm_address_t)
 168                 thread_to_be_cleaned_up->control_stack_start;
 169             next->next = new_freeable_stack;
 170             freeable_stack_count++;
 171         } else {
 172             struct freeable_stack *new_freeable_stack = 0;
 173             new_freeable_stack = (struct freeable_stack *)
 174                 os_validate(0, sizeof(struct freeable_stack));
 175             new_freeable_stack->next = NULL;
 176             new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 177             new_freeable_stack->stack = (os_vm_address_t)
 178                 thread_to_be_cleaned_up->control_stack_start;
 179             freeable_stack_queue = new_freeable_stack;
 180             freeable_stack_count++;
 181         }
 182         pthread_mutex_unlock(&freeable_stack_lock);
 183     }
 184 }
 185
 186 #define FREEABLE_STACK_QUEUE_SIZE 4
 187
 188 static void
 189 free_freeable_stacks() {
 190     if (freeable_stack_queue && (freeable_stack_count > FREEABLE_STACK_QUEUE_SIZE)) {
 191         struct freeable_stack* old;
 192         pthread_mutex_lock(&freeable_stack_lock);
 193         old = freeable_stack_queue;
 194         freeable_stack_queue = old->next;
 195         freeable_stack_count--;
 196         gc_assert(pthread_join(old->os_thread, NULL) == 0);
 197         FSHOW((stderr, "freeing thread %x stack\n", old->os_thread));
 198         os_invalidate(old->stack, THREAD_STRUCT_SIZE);
 199         os_invalidate((os_vm_address_t)old, sizeof(struct freeable_stack));
 200         pthread_mutex_unlock(&freeable_stack_lock);
 201     }
 202 }
 203
 204 #elif defined(CREATE_CLEANUP_THREAD)
 205 static void *
 206 cleanup_thread(void *arg)
 207 {
 208     struct freeable_stack *freeable = arg;
 209     pthread_t self = pthread_self();
 210
 211     FSHOW((stderr, "/cleaner thread(%p): joining %p\n",
 212            self, freeable->os_thread));
 213     gc_assert(pthread_join(freeable->os_thread, NULL) == 0);
 214     FSHOW((stderr, "/cleaner thread(%p): free stack %p\n",
 215            self, freeable->stack));
 216     os_invalidate(freeable->stack, THREAD_STRUCT_SIZE);
 217     free(freeable);
 218
 219     pthread_detach(self);
 220
 221     return NULL;
 222 }
 223
 224 static void
 225 create_cleanup_thread(struct thread *thread_to_be_cleaned_up)
 226 {
 227     pthread_t thread;
 228     int result;
 229
 230     if (thread_to_be_cleaned_up) {
 231         struct freeable_stack *freeable =
 232             malloc(sizeof(struct freeable_stack));
 233         gc_assert(freeable != NULL);
 234         freeable->os_thread = thread_to_be_cleaned_up->os_thread;
 235         freeable->stack =
 236             (os_vm_address_t) thread_to_be_cleaned_up->control_stack_start;
 237         result = pthread_create(&thread, NULL, cleanup_thread, freeable);
 238         gc_assert(result == 0);
 239         sched_yield();
 240     }
 241 }
 242
 243 #else
 244 static void
 245 free_thread_stack_later(struct thread *thread_to_be_cleaned_up)
 246 {
 247     struct freeable_stack *new_freeable_stack = 0;
 248     if (thread_to_be_cleaned_up) {
 249         new_freeable_stack = (struct freeable_stack *)
 250             os_validate(0, sizeof(struct freeable_stack));
 251         new_freeable_stack->os_thread = thread_to_be_cleaned_up->os_thread;
 252         new_freeable_stack->stack = (os_vm_address_t)
 253             thread_to_be_cleaned_up->control_stack_start;
 254     }
 255     new_freeable_stack = (struct freeable_stack *)
 256         swap_lispobjs((lispobj *)(void *)&freeable_stack,
 257                       (lispobj)new_freeable_stack);
 258     if (new_freeable_stack) {
 259         FSHOW((stderr,"/reaping %p\n", (void*) new_freeable_stack->os_thread));
 260         /* Under NPTL pthread_join really waits until the thread
 261          * exists and the stack can be safely freed. This is sadly not
 262          * mandated by the pthread spec. */
 263         gc_assert(pthread_join(new_freeable_stack->os_thread, NULL) == 0);
 264         os_invalidate(new_freeable_stack->stack, THREAD_STRUCT_SIZE);
 265         os_invalidate((os_vm_address_t) new_freeable_stack,
 266                       sizeof(struct freeable_stack));
 267     }
 268 }
 269 #endif
 270
 271 /* this is the first thing that runs in the child (which is why the
 272  * silly calling convention).  Basically it calls the user's requested
 273  * lisp function after doing arch_os_thread_init and whatever other
 274  * bookkeeping needs to be done
 275  */
 276 int
 277 new_thread_trampoline(struct thread *th)
 278 {
 279     lispobj function;
 280     int result, lock_ret;
 281
 282     FSHOW((stderr,"/creating thread %lu\n", thread_self()));
 283     function = th->no_tls_value_marker;
 284     th->no_tls_value_marker = NO_TLS_VALUE_MARKER_WIDETAG;
 285     if(arch_os_thread_init(th)==0) {
 286         /* FIXME: handle error */
 287         lose("arch_os_thread_init failed\n");
 288     }
 289
 290     th->os_thread=thread_self();
 291     protect_control_stack_guard_page(1);
 292     /* Since GC can only know about this thread from the all_threads
 293      * list and we're just adding this thread to it there is no danger
 294      * of deadlocking even with SIG_STOP_FOR_GC blocked (which it is
 295      * not). */
 296     lock_ret = pthread_mutex_lock(&all_threads_lock);
 297     gc_assert(lock_ret == 0);
 298     link_thread(th);
 299     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 300     gc_assert(lock_ret == 0);
 301
 302     result = funcall0(function);
 303
 304     /* Block GC */
 305     block_blockable_signals();
 306     th->state=STATE_DEAD;
 307
 308     /* SIG_STOP_FOR_GC is blocked and GC might be waiting for this
 309      * thread, but since we are already dead it won't wait long. */
 310     lock_ret = pthread_mutex_lock(&all_threads_lock);
 311     gc_assert(lock_ret == 0);
 312
 313     gc_alloc_update_page_tables(0, &th->alloc_region);
 314     unlink_thread(th);
 315     pthread_mutex_unlock(&all_threads_lock);
 316     gc_assert(lock_ret == 0);
 317
 318     if(th->tls_cookie>=0) arch_os_thread_cleanup(th);
 319     os_invalidate((os_vm_address_t)th->interrupt_data,
 320                   (sizeof (struct interrupt_data)));
 321
 322 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 323     FSHOW((stderr, "Deallocating mach port %x\n", THREAD_STRUCT_TO_EXCEPTION_PORT(th)));
 324     mach_port_move_member(mach_task_self(),
 325                           THREAD_STRUCT_TO_EXCEPTION_PORT(th),
 326                           MACH_PORT_NULL);
 327     mach_port_deallocate(mach_task_self(),
 328                          THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 329     mach_port_destroy(mach_task_self(),
 330                       THREAD_STRUCT_TO_EXCEPTION_PORT(th));
 331 #endif
 332
 333 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 334     queue_freeable_thread_stack(th);
 335 #elif defined(CREATE_CLEANUP_THREAD)
 336     create_cleanup_thread(th);
 337 #else
 338     free_thread_stack_later(th);
 339 #endif
 340
 341     FSHOW((stderr,"/exiting thread %p\n", thread_self()));
 342     return result;
 343 }
 344
 345 #endif /* LISP_FEATURE_SB_THREAD */
 346
 347 static void
 348 free_thread_struct(struct thread *th)
 349 {
 350     if (th->interrupt_data)
 351         os_invalidate((os_vm_address_t) th->interrupt_data,
 352                       (sizeof (struct interrupt_data)));
 353     os_invalidate((os_vm_address_t) th->control_stack_start,
 354                   THREAD_STRUCT_SIZE);
 355 }
 356
 357 /* this is called from any other thread to create the new one, and
 358  * initialize all parts of it that can be initialized from another
 359  * thread
 360  */
 361
 362 static struct thread *
 363 create_thread_struct(lispobj initial_function) {
 364     union per_thread_data *per_thread;
 365     struct thread *th=0;        /*  subdue gcc */
 366     void *spaces=0;
 367 #ifdef LISP_FEATURE_SB_THREAD
 368     int i;
 369 #endif
 370
 371 #ifdef CREATE_CLEANUP_THREAD
 372     /* Give a chance for cleanup threads to run. */
 373     sched_yield();
 374 #endif
 375     /* may as well allocate all the spaces at once: it saves us from
 376      * having to decide what to do if only some of the allocations
 377      * succeed */
 378     spaces=os_validate(0, THREAD_STRUCT_SIZE);
 379     if(!spaces)
 380          return NULL;
 381     per_thread=(union per_thread_data *)
 382         (spaces+
 383          THREAD_CONTROL_STACK_SIZE+
 384          BINDING_STACK_SIZE+
 385          ALIEN_STACK_SIZE);
 386
 387 #ifdef LISP_FEATURE_SB_THREAD
 388     for(i = 0; i < (dynamic_values_bytes / sizeof(lispobj)); i++)
 389         per_thread->dynamic_values[i] = NO_TLS_VALUE_MARKER_WIDETAG;
 390     if (all_threads == 0) {
 391         if(SymbolValue(FREE_TLS_INDEX,0)==UNBOUND_MARKER_WIDETAG) {
 392             SetSymbolValue
 393                 (FREE_TLS_INDEX,
 394                  /* FIXME: should be MAX_INTERRUPTS -1 ? */
 395                  make_fixnum(MAX_INTERRUPTS+
 396                              sizeof(struct thread)/sizeof(lispobj)),
 397                  0);
 398             SetSymbolValue(TLS_INDEX_LOCK,make_fixnum(0),0);
 399         }
 400 #define STATIC_TLS_INIT(sym,field) \
 401   ((struct symbol *)(sym-OTHER_POINTER_LOWTAG))->tls_index= \
 402   make_fixnum(THREAD_SLOT_OFFSET_WORDS(field))
 403
 404         STATIC_TLS_INIT(BINDING_STACK_START,binding_stack_start);
 405         STATIC_TLS_INIT(BINDING_STACK_POINTER,binding_stack_pointer);
 406         STATIC_TLS_INIT(CONTROL_STACK_START,control_stack_start);
 407         STATIC_TLS_INIT(CONTROL_STACK_END,control_stack_end);
 408         STATIC_TLS_INIT(ALIEN_STACK,alien_stack_pointer);
 409 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 410         STATIC_TLS_INIT(PSEUDO_ATOMIC_BITS,pseudo_atomic_bits);
 411 #endif
 412 #undef STATIC_TLS_INIT
 413     }
 414 #endif
 415
 416     th=&per_thread->thread;
 417     th->control_stack_start = spaces;
 418     th->binding_stack_start=
 419         (lispobj*)((void*)th->control_stack_start+THREAD_CONTROL_STACK_SIZE);
 420     th->control_stack_end = th->binding_stack_start;
 421     th->alien_stack_start=
 422         (lispobj*)((void*)th->binding_stack_start+BINDING_STACK_SIZE);
 423     th->binding_stack_pointer=th->binding_stack_start;
 424     th->this=th;
 425     th->os_thread=0;
 426     th->state=STATE_RUNNING;
 427 #ifdef LISP_FEATURE_STACK_GROWS_DOWNWARD_NOT_UPWARD
 428     th->alien_stack_pointer=((void *)th->alien_stack_start
 429                              + ALIEN_STACK_SIZE-N_WORD_BYTES);
 430 #else
 431     th->alien_stack_pointer=((void *)th->alien_stack_start);
 432 #endif
 433 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 434     th->pseudo_atomic_bits=0;
 435 #endif
 436 #ifdef LISP_FEATURE_GENCGC
 437     gc_set_region_empty(&th->alloc_region);
 438 #endif
 439
 440 #ifndef LISP_FEATURE_SB_THREAD
 441     /* the tls-points-into-struct-thread trick is only good for threaded
 442      * sbcl, because unithread sbcl doesn't have tls.  So, we copy the
 443      * appropriate values from struct thread here, and make sure that
 444      * we use the appropriate SymbolValue macros to access any of the
 445      * variable quantities from the C runtime.  It's not quite OAOOM,
 446      * it just feels like it */
 447     SetSymbolValue(BINDING_STACK_START,(lispobj)th->binding_stack_start,th);
 448     SetSymbolValue(CONTROL_STACK_START,(lispobj)th->control_stack_start,th);
 449     SetSymbolValue(CONTROL_STACK_END,(lispobj)th->control_stack_end,th);
 450 #if defined(LISP_FEATURE_X86) || defined (LISP_FEATURE_X86_64)
 451     SetSymbolValue(BINDING_STACK_POINTER,(lispobj)th->binding_stack_pointer,th);
 452     SetSymbolValue(ALIEN_STACK,(lispobj)th->alien_stack_pointer,th);
 453     SetSymbolValue(PSEUDO_ATOMIC_BITS,(lispobj)th->pseudo_atomic_bits,th);
 454 #else
 455     current_binding_stack_pointer=th->binding_stack_pointer;
 456     current_control_stack_pointer=th->control_stack_start;
 457 #endif
 458 #endif
 459     bind_variable(CURRENT_CATCH_BLOCK,make_fixnum(0),th);
 460     bind_variable(CURRENT_UNWIND_PROTECT_BLOCK,make_fixnum(0),th);
 461     bind_variable(FREE_INTERRUPT_CONTEXT_INDEX,make_fixnum(0),th);
 462     bind_variable(INTERRUPT_PENDING, NIL,th);
 463     bind_variable(INTERRUPTS_ENABLED,T,th);
 464     bind_variable(GC_PENDING,NIL,th);
 465 #ifdef LISP_FEATURE_SB_THREAD
 466     bind_variable(STOP_FOR_GC_PENDING,NIL,th);
 467 #endif
 468
 469     th->interrupt_data = (struct interrupt_data *)
 470         os_validate(0,(sizeof (struct interrupt_data)));
 471     if (!th->interrupt_data) {
 472         free_thread_struct(th);
 473         return 0;
 474     }
 475     th->interrupt_data->pending_handler = 0;
 476     th->no_tls_value_marker=initial_function;
 477
 478     th->stepping = NIL;
 479     return th;
 480 }
 481
 482 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 483 mach_port_t setup_mach_exception_handling_thread();
 484 kern_return_t mach_thread_init(mach_port_t thread_exception_port);
 485
 486 #endif
 487
 488 void create_initial_thread(lispobj initial_function) {
 489     struct thread *th=create_thread_struct(initial_function);
 490     if(th) {
 491 #ifdef LISP_FEATURE_MACH_EXCEPTION_HANDLER
 492         kern_return_t ret;
 493
 494         setup_mach_exception_handling_thread();
 495 #endif
 496         initial_thread_trampoline(th); /* no return */
 497     } else lose("can't create initial thread\n");
 498 }
 499
 500 #ifdef LISP_FEATURE_SB_THREAD
 501
 502 #ifndef __USE_XOPEN2K
 503 extern int pthread_attr_setstack (pthread_attr_t *__attr, void *__stackaddr,
 504                                   size_t __stacksize);
 505 #endif
 506
 507 boolean create_os_thread(struct thread *th,os_thread_t *kid_tid)
 508 {
 509     /* The new thread inherits the restrictive signal mask set here,
 510      * and enables signals again when it is set up properly. */
 511     pthread_attr_t attr;
 512     sigset_t newset,oldset;
 513     boolean r=1;
 514     int retcode, initcode, sizecode, addrcode;
 515
 516     FSHOW_SIGNAL((stderr,"/create_os_thread: creating new thread\n"));
 517
 518 #ifdef LOCK_CREATE_THREAD
 519     retcode = pthread_mutex_lock(&create_thread_lock);
 520     gc_assert(retcode == 0);
 521     FSHOW_SIGNAL((stderr,"/create_os_thread: got lock\n"));
 522 #endif
 523     sigemptyset(&newset);
 524     /* Blocking deferrable signals is enough, no need to block
 525      * SIG_STOP_FOR_GC because the child process is not linked onto
 526      * all_threads until it's ready. */
 527     sigaddset_deferrable(&newset);
 528     thread_sigmask(SIG_BLOCK, &newset, &oldset);
 529
 530 #if defined(LISP_FEATURE_DARWIN)
 531 #define CONTROL_STACK_ADJUST 8192 /* darwin wants page-aligned stacks */
 532 #else
 533 #define CONTROL_STACK_ADJUST 16
 534 #endif
 535
 536     if((initcode = pthread_attr_init(&attr)) ||
 537        /* FIXME: why do we even have this in the first place? */
 538        (pthread_attr_setstack(&attr,th->control_stack_start,
 539                               THREAD_CONTROL_STACK_SIZE-CONTROL_STACK_ADJUST)) ||
 540 #undef CONTROL_STACK_ADJUST
 541        (retcode = pthread_create
 542         (kid_tid,&attr,(void *(*)(void *))new_thread_trampoline,th))) {
 543         FSHOW_SIGNAL((stderr, "init, size, addr = %d, %d, %d\n", initcode, sizecode, addrcode));
 544         FSHOW_SIGNAL((stderr, printf("pthread_create returned %d, errno %d\n", retcode, errno)));
 545         FSHOW_SIGNAL((stderr, "wanted stack size %d, min stack size %d\n",
 546                       THREAD_CONTROL_STACK_SIZE-16, PTHREAD_STACK_MIN));
 547         if(retcode < 0) {
 548             perror("create_os_thread");
 549         }
 550         r=0;
 551     }
 552
 553 #ifdef QUEUE_FREEABLE_THREAD_STACKS
 554     free_freeable_stacks();
 555 #endif
 556     thread_sigmask(SIG_SETMASK,&oldset,0);
 557 #ifdef LOCK_CREATE_THREAD
 558     retcode = pthread_mutex_unlock(&create_thread_lock);
 559     gc_assert(retcode == 0);
 560     FSHOW_SIGNAL((stderr,"/create_os_thread: released lock\n"));
 561 #endif
 562     return r;
 563 }
 564
 565 os_thread_t create_thread(lispobj initial_function) {
 566     struct thread *th;
 567     os_thread_t kid_tid;
 568
 569     /* Assuming that a fresh thread struct has no lisp objects in it,
 570      * linking it to all_threads can be left to the thread itself
 571      * without fear of gc lossage. initial_function violates this
 572      * assumption and must stay pinned until the child starts up. */
 573     th = create_thread_struct(initial_function);
 574     if(th==0) return 0;
 575
 576     if (create_os_thread(th,&kid_tid)) {
 577         return kid_tid;
 578     } else {
 579         free_thread_struct(th);
 580         return 0;
 581     }
 582 }
 583
 584 /* Send the signo to os_thread, retry if the rt signal queue is
 585  * full. */
 586 int
 587 kill_thread_safely(os_thread_t os_thread, int signo)
 588 {
 589     int r;
 590     /* The man page does not mention EAGAIN as a valid return value
 591      * for either pthread_kill or kill. But that's theory, this is
 592      * practice. By waiting here we assume that the delivery of this
 593      * signal is not necessary for the delivery of the signals in the
 594      * queue. In other words, we _assume_ there are no deadlocks. */
 595     while ((r=pthread_kill(os_thread,signo))==EAGAIN) {
 596         /* wait a bit then try again in the hope of the rt signal
 597          * queue not being full */
 598         FSHOW_SIGNAL((stderr,"/rt signal queue full\n"));
 599         /* FIXME: some kind of backoff (random, exponential) would be
 600          * nice. */
 601         sleep(1);
 602     }
 603     return r;
 604 }
 605
 606 int signal_interrupt_thread(os_thread_t os_thread)
 607 {
 608     int status = kill_thread_safely(os_thread, SIG_INTERRUPT_THREAD);
 609     if (status == 0) {
 610         return 0;
 611     } else if (status == ESRCH) {
 612         return -1;
 613     } else {
 614         lose("cannot send SIG_INTERRUPT_THREAD to thread=%lu: %d, %s\n",
 615              os_thread, status, strerror(status));
 616     }
 617 }
 618
 619 /* stopping the world is a two-stage process.  From this thread we signal
 620  * all the others with SIG_STOP_FOR_GC.  The handler for this signal does
 621  * the usual pseudo-atomic checks (we don't want to stop a thread while
 622  * it's in the middle of allocation) then waits for another SIG_STOP_FOR_GC.
 623  */
 624
 625 /* To avoid deadlocks when gc stops the world all clients of each
 626  * mutex must enable or disable SIG_STOP_FOR_GC for the duration of
 627  * holding the lock, but they must agree on which. */
 628 void gc_stop_the_world()
 629 {
 630     struct thread *p,*th=arch_os_get_current_thread();
 631     int status, lock_ret;
 632 #ifdef LOCK_CREATE_THREAD
 633     /* KLUDGE: Stopping the thread during pthread_create() causes deadlock
 634      * on FreeBSD. */
 635     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on create_thread_lock, thread=%lu\n",
 636                   th->os_thread));
 637     lock_ret = pthread_mutex_lock(&create_thread_lock);
 638     gc_assert(lock_ret == 0);
 639     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got create_thread_lock, thread=%lu\n",
 640                   th->os_thread));
 641 #endif
 642     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:waiting on lock, thread=%lu\n",
 643                   th->os_thread));
 644     /* keep threads from starting while the world is stopped. */
 645     lock_ret = pthread_mutex_lock(&all_threads_lock);      \
 646     gc_assert(lock_ret == 0);
 647
 648     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:got lock, thread=%lu\n",
 649                   th->os_thread));
 650     /* stop all other threads by sending them SIG_STOP_FOR_GC */
 651     for(p=all_threads; p; p=p->next) {
 652         gc_assert(p->os_thread != 0);
 653         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: p->state: %x\n", p->state));
 654         if((p!=th) && ((p->state==STATE_RUNNING))) {
 655             FSHOW_SIGNAL((stderr,"/gc_stop_the_world: suspending %x, os_thread %x\n",
 656                           p, p->os_thread));
 657             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 658             if (status==ESRCH) {
 659                 /* This thread has exited. */
 660                 gc_assert(p->state==STATE_DEAD);
 661             } else if (status) {
 662                 lose("cannot send suspend thread=%lu: %d, %s\n",
 663                      p->os_thread,status,strerror(status));
 664             }
 665         }
 666     }
 667     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:signals sent\n"));
 668     /* wait for the running threads to stop or finish */
 669     for(p=all_threads;p;) {
 670         FSHOW_SIGNAL((stderr,"/gc_stop_the_world: th: %p, p: %p\n", th, p));
 671         if((p!=th) && (p->state==STATE_RUNNING)) {
 672             sched_yield();
 673         } else {
 674             p=p->next;
 675         }
 676     }
 677     FSHOW_SIGNAL((stderr,"/gc_stop_the_world:end\n"));
 678 }
 679
 680 void gc_start_the_world()
 681 {
 682     struct thread *p,*th=arch_os_get_current_thread();
 683     int status, lock_ret;
 684     /* if a resumed thread creates a new thread before we're done with
 685      * this loop, the new thread will get consed on the front of
 686      * all_threads, but it won't have been stopped so won't need
 687      * restarting */
 688     FSHOW_SIGNAL((stderr,"/gc_start_the_world:begin\n"));
 689     for(p=all_threads;p;p=p->next) {
 690         gc_assert(p->os_thread!=0);
 691         if((p!=th) && (p->state!=STATE_DEAD)) {
 692             if(p->state!=STATE_SUSPENDED) {
 693                 lose("gc_start_the_world: wrong thread state is %d\n",
 694                      fixnum_value(p->state));
 695             }
 696             FSHOW_SIGNAL((stderr, "/gc_start_the_world: resuming %lu\n",
 697                           p->os_thread));
 698             p->state=STATE_RUNNING;
 699
 700 #if defined(SIG_RESUME_FROM_GC)
 701             status=kill_thread_safely(p->os_thread,SIG_RESUME_FROM_GC);
 702 #else
 703             status=kill_thread_safely(p->os_thread,SIG_STOP_FOR_GC);
 704 #endif
 705             if (status) {
 706                 lose("cannot resume thread=%lu: %d, %s\n",
 707                      p->os_thread,status,strerror(status));
 708             }
 709         }
 710     }
 711     /* If we waited here until all threads leave STATE_SUSPENDED, then
 712      * SIG_STOP_FOR_GC wouldn't need to be a rt signal. That has some
 713      * performance implications, but does away with the 'rt signal
 714      * queue full' problem. */
 715
 716     lock_ret = pthread_mutex_unlock(&all_threads_lock);
 717     gc_assert(lock_ret == 0);
 718 #ifdef LOCK_CREATE_THREAD
 719     lock_ret = pthread_mutex_unlock(&create_thread_lock);
 720     gc_assert(lock_ret == 0);
 721 #endif
 722
 723     FSHOW_SIGNAL((stderr,"/gc_start_the_world:end\n"));
 724 }
 725 #endif